app.py

import streamlit as st
import os
import base64
import io
from PIL import Image
from pydub import AudioSegment
import IPython
import soundfile as sf
import requests
import pandas as pd  # If you're working with DataFrames
import matplotlib.figure  # If you're using matplotlib figures
import numpy as np

# For Altair charts
import altair as alt

# For Bokeh charts
from bokeh.models import Plot

# For Plotly charts
import plotly.express as px

# For Pydeck charts
import pydeck as pdk


import time  
from transformers import load_tool, Agent
import torch

class ToolLoader:
    def __init__(self, tool_names):
        self.tools = self.load_tools(tool_names)

    def load_tools(self, tool_names):
        loaded_tools = []
        for tool_name in tool_names:
            try:
                tool = load_tool(tool_name)
                loaded_tools.append(tool)
            except Exception as e:
                log_response(f"Error loading tool '{tool_name}': {e}")
        return loaded_tools

class CustomHfAgent(Agent):
    def __init__(self, url_endpoint, token, chat_prompt_template=None, run_prompt_template=None, additional_tools=None, input_params=None):
        super().__init__(
            chat_prompt_template=chat_prompt_template,
            run_prompt_template=run_prompt_template,
            additional_tools=additional_tools,
        )
        self.url_endpoint = url_endpoint
        self.token = token
        self.input_params = input_params

    def generate_one(self, prompt, stop):
        headers = {"Authorization": self.token}
        max_new_tokens = self.input_params.get("max_new_tokens", 192)
        parameters = {"max_new_tokens": max_new_tokens, "return_full_text": False, "stop": stop, "padding": True, "truncation": True}
        inputs = {
            "inputs": prompt,
            "parameters": parameters,
        }
        response = requests.post(self.url_endpoint, json=inputs, headers=headers)

        if response.status_code == 429:
            log_response("Getting rate-limited, waiting a tiny bit before trying again.")
            time.sleep(1)
            return self._generate_one(prompt)
        elif response.status_code != 200:
            raise ValueError(f"Errors {inputs} {response.status_code}: {response.json()}")
        log_response(response)
        result = response.json()[0]["generated_text"]
        for stop_seq in stop:
            if result.endswith(stop_seq):
                return result[: -len(stop_seq)]
        return result

def handle_submission(user_message, selected_tools, url_endpoint):
    
    log_response("User input \n {}".format(user_message))
    log_response("selected_tools \n {}".format(selected_tools))
    log_response("url_endpoint \n {}".format(url_endpoint))
    
    agent = CustomHfAgent(
        url_endpoint=url_endpoint,
        token=os.environ['HF_token'],
        additional_tools=selected_tools,
        input_params={"max_new_tokens": 192},
    )

    response = agent.run(user_message)

    log_response("Agent Response\n {}".format(response))

    return response

# Declare global variable
global log_enabled
log_enabled = False

def log_response(response):
    if log_enabled:
        with st.chat_message("ai"):
            st.markdown("Agent Response\n {}".format(response))
    print(response)

    
# Define the tool names to load
tool_names = [
    "Chris4K/random-character-tool",
    "Chris4K/text-generation-tool",
    "Chris4K/sentiment-tool", 
    "Chris4K/token-counter-tool", 
    "Chris4K/most-downloaded-model",
    "Chris4K/rag-tool",
    "Chris4K/word-counter-tool",
    "Chris4K/sentence-counter-tool",
    "Chris4K/EmojifyTextTool",
    "Chris4K/NamedEntityRecognitionTool",
    "Chris4K/TextDownloadTool",
    "Chris4K/source-code-retriever-tool",
    "Chris4K/text-to-image",
    "Chris4K/text-to-video",
    "Chris4K/image-transformation",
    "Chris4K/latent-upscaler-tool"
    # More cool tools to come
]

# Create tool loader instance
tool_loader = ToolLoader(tool_names)

st.title("Hugging Face Agent and tools")

## LB https://huggingface.co./spaces/qiantong-xu/toolbench-leaderboard

st.markdown("Welcome to the Hugging Face Agent and Tools app! This app allows you to interact with various tools using the Hugging Face API.")
    
# Create a page with tabs
tabs = st.tabs(["Chat", "URL and Tools", "User Description", "Developers"])

# Tab 1: Chat
with tabs[0]:
    
    # Code for URL and Tools checkboxes 

    # Examples for the user perspective
    st.markdown("### Examples:")
    st.markdown("1. **Generate a Random Character**:")
    st.markdown("   - Choose the desired URL and the 'Random Character Tool'.")
    
    st.markdown("2. **Sentiment Analysis**:")
    st.markdown("   - Choose the desired URL and the 'Sentiment Analysis Tool'.")
    st.markdown("   - Sample: What is the sentiment for \"Hello, I am happy\"?")
    
    st.markdown("3. **Word Count**:")
    st.markdown("   - Choose the desired URL and the 'Word Counter Tool'.")
    st.markdown("   - Sample: Count the words in \"Hello, I am Christof\".")
 
        
# Tab 2: URL and Tools
with tabs[1]:
 
    # Code for URL and Tools checkboxes 

    # Examples for the user perspective
    st.markdown("### Examples:")
    st.markdown("1. **Generate a Random Character**:")
    st.markdown("   - Choose the desired URL and the 'Random Character Tool'.")
    
    st.markdown("2. **Sentiment Analysis**:")
    st.markdown("   - Choose the desired URL and the 'Sentiment Analysis Tool'.")
    st.markdown("   - Sample: What is the sentiment for \"Hello, I am happy\"?")
    
    st.markdown("3. **Word Count**:")
    st.markdown("   - Choose the desired URL and the 'Word Counter Tool'.")
    st.markdown("   - Sample: Count the words in \"Hello, I am Christof\".")

    # Add a dropdown for selecting the inference URL
    url_endpoint = st.selectbox("Select Inference URL", [
        "https://api-inference.huggingface.co/models/bigcode/starcoder",
        "https://api-inference.huggingface.co/models/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
        "https://api-inference.huggingface.co/models/gpt2"
    ])

        # Add a checkbox for enabling logging
    log_enabled = st.checkbox("Enable Logging")
    
    tool_checkboxes = [st.checkbox(f"{tool.name} --- {tool.description} ") for tool in tool_loader.tools]

# Tab 3: User Description
with tabs[2]:
 
    # User description content and tool descriptions
    # Add a section for the app's description
    st.markdown(''' 
                # Hugging Face Agent and Tools App
                
                ## Description
                Welcome to the Hugging Face Agent and Tools app! This app provides an interactive interface for utilizing various tools through the Hugging Face API. You can choose an inference URL and select from a variety of tools to perform different tasks.
                
                ## Examples
                1. **Generate a Random Character**:
                   - Choose the desired URL and the 'Random Character Tool'.
                   
                2. **Sentiment Analysis**:
                   - Choose the desired URL and the 'Sentiment Analysis Tool'.
                   - Sample: What is the sentiment for "Hello, I am happy"?
                
                3. **Word Count**:
                   - Choose the desired URL and the 'Word Counter Tool'.
                   - Sample: Count the words in "Hello, I am Christof".
                
                ## Tools
                To interact with the tools, expand the section below to see tool descriptions and select the tools you want to use.
                
                <details>
                <summary>Expand to see tool descriptions</summary>
                
                ### Tool Descriptions
                - **random-character-tool:** Generates a random character.
                - **text-generation-tool:** Generates text based on a prompt.
                - **sentiment-tool:** Analyzes the sentiment of a given text.
                - **token-counter-tool:** Counts the tokens in a text.
                - **most-downloaded-model:** Provides information about the most downloaded model.
                - **rag-tool:** Utilizes Retrieval-Augmented Generation (RAG) for text generation.
                - **word-counter-tool:** Counts the words in a text.
                - **sentence-counter-tool:** Counts the sentences in a text.
                - **EmojifyTextTool:** Emojifies the given text.
                - **NamedEntityRecognitionTool:** Identifies named entities in a text.
                - **TextDownloadTool:** Downloads text from a given URL.
                - **source-code-retriever-tool:** Retrieves source code from a given URL.
                - **text-to-image:** Generates an image from text.
                - **text-to-video:** Generates a video from text.
                - **image-transformation:** Applies transformations to images.
                - **latent-upscaler-tool:** Upscales images using latent space.
                
                </details>
                
                ## Usage
                1. Choose the desired inference URL from the dropdown.
                2. Expand the tool selection section and choose the tools you want to use.
                3. Enter a message in the chat input to interact with the Hugging Face Agent.
                4. View the assistant's responses, which may include images, audio, text, or other visualizations based on the selected tools.
                
                Feel free to explore and experiment with different tools to achieve various tasks!

            ''')

# Tab 4: Developers
with tabs[3]:
         
    # Developer-related content
    st.markdown(''' 

        # Hugging Face Agent and Tools Code Overview
        
        ## Overview
        The provided Python code implements an interactive Streamlit web application that allows users to interact with various tools through the Hugging Face API. The app integrates Hugging Face models and tools, enabling users to perform tasks such as text generation, sentiment analysis, and more.
        
        ## Imports
        The code imports several external libraries and modules, including:
        - `streamlit`: For building the web application.
        - `os`: For interacting with the operating system.
        - `base64`, `io`, `Image` (from `PIL`), `AudioSegment` (from `pydub`), `IPython`, `sf`: For handling images and audio.
        - `requests`: For making HTTP requests.
        - `pandas`: For working with DataFrames.
        - `matplotlib.figure`, `numpy`: For visualization.
        - `altair`, `Plot` (from `bokeh.models`), `px` (from `plotly.express`), `pdk` (from `pydeck`): For different charting libraries.
        - `time`: For handling time-related operations.
        - `transformers`: For loading tools and agents.
        
        ## ToolLoader Class
        The `ToolLoader` class is responsible for loading tools based on their names. It has methods to load tools from a list of tool names and handles potential errors during loading.
        
        ## CustomHfAgent Class
        The `CustomHfAgent` class extends the base `Agent` class from the `transformers` module. It is designed to interact with a remote inference API and includes methods for generating text based on a given prompt.
        
        ## Tool Loading and Customization
        - Tool names are defined in the `tool_names` list.
        - The `ToolLoader` instance (`tool_loader`) loads tools based on the provided names.
        - The `CustomHfAgent` instance (`agent`) is created with a specified URL endpoint, token, and additional tools.
        - New tools can be added by appending their names to the `tool_names` list.
        
        ## Streamlit App
        The Streamlit app is structured as follows:
        1. Tool selection dropdown for choosing the inference URL.
        2. An expander for displaying tool descriptions.
        3. An expander for selecting tools.
        4. Examples and instructions for the user.
        5. A chat interface for user interactions.
        6. Handling of user inputs, tool selection, and agent responses.
        
        ## Handling of Responses
        The code handles various types of responses from the agent, including images, audio, text, DataFrames, and charts. The responses are displayed in the Streamlit app based on their types.
        
        ## How to Run
        1. Install required dependencies with `pip install -r requirements.txt`.
        2. Run the app with `streamlit run <filename.py>`.
        
        ## Notes
        - The code emphasizes customization and extensibility, allowing developers to easily add new tools and interact with the Hugging Face API.
        - Ensure proper configuration, such as setting the Hugging Face token as an environment variable.

      ''')
    
# Chat code (user input, agent responses, etc.)
if "messages" not in st.session_state:
    st.session_state.messages = []

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.markdown(message["content"])

with st.chat_message("assistant"):
    st.markdown("Hello there! How can I assist you today?")

if user_message := st.chat_input("Enter message"):
    st.chat_message("user").markdown(user_message)
    st.session_state.messages.append({"role": "user", "content": user_message})

    selected_tools = [tool_loader.tools[idx] for idx, checkbox in enumerate(tool_checkboxes) if checkbox]
    # Handle submission with the selected inference URL
    response = handle_submission(user_message, selected_tools, url_endpoint)

    with st.chat_message("assistant"):
        if response is None:
            st.warning("The agent's response is None. Please try again. Generate an image of a flying horse.")
        elif isinstance(response, Image.Image):
            st.image(response)
        elif isinstance(response, AudioSegment):
            st.audio(response)
        elif isinstance(response, int):
            st.markdown(response)
        elif isinstance(response, str):
            if "emojified_text" in response:
                st.markdown(f"{response['emojified_text']}")
            else:
                st.markdown(response)
        elif isinstance(response, list):
            for item in response:
                st.markdown(item)  # Assuming the list contains strings
        elif isinstance(response, pd.DataFrame):
            st.dataframe(response)
        elif isinstance(response, pd.Series):
            st.table(response.iloc[0:10])
        elif isinstance(response, dict):
            st.json(response)
        elif isinstance(response, st.graphics_altair.AltairChart):
            st.altair_chart(response)
        elif isinstance(response, st.graphics_bokeh.BokehChart):
            st.bokeh_chart(response)
        elif isinstance(response, st.graphics_graphviz.GraphvizChart):
            st.graphviz_chart(response)
        elif isinstance(response, st.graphics_plotly.PlotlyChart):
            st.plotly_chart(response)
        elif isinstance(response, st.graphics_pydeck.PydeckChart):
            st.pydeck_chart(response)
        elif isinstance(response, matplotlib.figure.Figure):
            st.pyplot(response)
        elif isinstance(response, streamlit.graphics_vega_lite.VegaLiteChart):
            st.vega_lite_chart(response)
        else:
            st.warning("Unrecognized response type. Please try again. e.g. Generate an image of a flying horse.")

    st.session_state.messages.append({"role": "assistant", "content": response})